- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This work looks into sampling uncertainty in steering vectors, which are usually constructed by averaging hidden-state differences between examples of two opposite behaviors. The paper proposes a method called SteerCert that constructs a confidence region around the average vector, and supports a claim only when the claim holds across the region. Empirically the paper finds that A/B answer-position explains much of this variation. The paper also shows that stratified sampling reduces such variance.
Critical strengths
S1. The variability in steering vector construction and the sampling analysis is interesting.
Critical weaknesses
W1. The spherical region may be overly conservative. Sampling error in the estimated steering vector can differ substantially across directions, but a sphere assumes the same error magnitude in every direction. Requiring a claim to hold throughout the sphere may therefore cause false positive abstentions. In the paper there are only 50 of 1200 evaluations receive positive guarantee, which might result from overly conservative confidence region.
W2. The experiment assesses whether an infinitesimally small intervention increases the probability of an A/B answer. However, they do not test what happens at a more practical setting where the intervention strengths that are actually used for steering or when the model generates its own answers.
Questions for authors
Q1. The certificate rate is quite low (50 out of 1200). What are the false positive and false negative rates if you consider the actual behavioral effect as the ground truth label?
Q2. Would it be possible (and helpful) to extend this work to covariance-aware confidence regions, given the complexity of the distribution in the latent space? I wonder if this might help improve the certificate rate.
2 comments
W2 and Q1: Finite-strength responses and false-positive/negative rates
In the current manuscript, Sec. 5.2 shows that answer-position stratification improves response repeatability at finite strength for myopic reward. We now complement this result with additional experiments connecting local support to finite responses across the original 1,200 evaluations at three nonzero strengths. Most band-supported cases show positive answer-probability changes at each tested strength, with agreement declining at larger strengths.
W2: “they do not test what happens at a more practical setting where the intervention strengths that are actually used for steering”
Sec. 5.2 of the current manuscript evaluates finite-strength interventions across reconstructed steering directions. For myopic reward, answer-position stratification yields 0.067 times the response variance of ordinary sampling in Llama and 0.425 times in Gemma (, ), at the same target and sample size.
W2: “when the model generates its own answers”
Our behavioral evaluation uses answer-letter probabilities rather than scores of freely generated text. This lets us measure construction variability and compare local signs with finite-strength effects on the same response measure.
Q1: “What are the false positive and false negative rates if you consider the actual behavioral effect as the ground truth label?”
For the additional analysis, we kept support selections fixed on all 1,200 evaluations and intervened along the normalized file-mean directions at strengths , where and $\kappa\in \lbrace 0.25,0.5,1 \rbrace $. The reference norm and absolute strength coefficients are the same as in Sec. 5.2. We use positive coefficients here because the evaluated claims concern positive local responses. The endpoint is the change in the behavior-matching answer letter’s full-vocabulary probability.
Of the 788 band-supported evaluations, 784, 758, and 691 had positive finite responses. Across all 788, median changes were +3.1, +5.9, and +5.7 percentage points, respectively. All 50 ball-supported evaluations had positive responses at all three strengths.
For the requested confusion matrix, withholding is treated as a negative classification, without asserting a negative effect. FPR is the fraction of nonpositive finite responses that were supported; FNR is the fraction of positive finite responses that were withheld.
| Ball FPR | Ball FNR | Projection-band FPR | Projection-band FNR | |
|---|---|---|---|---|
| 0.25 | 0.00% | 94.76% | 1.63% | 17.82% |
| 0.50 | 0.00% | 94.61% | 11.03% | 18.32% |
| 1.00 | 0.00% | 94.25% | 29.39% | 20.57% |
Among the 788 band-supported evaluations, 0.5%, 3.8%, and 12.3% had nonpositive finite responses at the three strengths, respectively. At the smallest tested strength, the band reduces FNR from 94.76% to 17.82%, with 1.63% FPR. These are descriptive rates on the fixed panel. Matched-count center selection gave 784, 761, and 696 positive responses out of 788; we found no consistent intervention-selection advantage on this panel.
At the largest strength, 97 supported evaluations had nonpositive finite responses despite positive local derivatives along the same file-mean directions. This measures the limit of extrapolating local signs to larger interventions, rather than false support for a population local sign.
Together, these results clarify both the gains in local-response support and the limits of extrapolating that support to finite-strength effects. We will revise Secs. 5.3–5.4 and the corresponding contribution, conclusion, and limitations; the appendix will provide calibration settings, effect distributions, and matched-count comparisons.
Thank you for the constructive suggestions on covariance-aware inference and finite-strength evaluation. We have added analyses addressing both points.
W1 and Q2: Conservatism and covariance-aware inference
A covariance-aware projection band recovers many correct local-response claims withheld by the ball in finite-source validation.
Following your suggestion, we compared the original confidence ball with a covariance-aware simultaneous projection band for the same predeclared local-response claims, using a multiplier bootstrap. Support increases from 50 to 788 of the original 1,200 evaluations, using the same construction data and nominal error allocation. Each simultaneous family contains 100 inputs for one behavior/layer combination.
The ball’s multiplier-bootstrap calibration allows for anisotropic sampling errors. Its spherical geometry can still be conservative for fixed local-response claims. This conservatism may be useful in steering analyses where limiting false support takes priority over recall.
These results suggest choosing the confidence construction to match the claim of interest. Both constructions apply the criterion in Sec. 3: a claim is statistically supported when it holds throughout the chosen confidence set. The projection band targets the 100 declared local-response claims using their joint uncertainty. The ball provides a common uncertainty set for assessing directional precision and local-response claims together, which can be useful when a steering direction is reused across inputs or analyses. The ball-based directional verdicts are unchanged.
We evaluated coverage and recall on expanded finite sources for sycophancy and hallucination. Across 2,000 reconstructions at n=1000:
Recall measures support among positive local signs at the known finite-source means. Coverage is averaged over six 100-input families; brackets give 95% bootstrap intervals conditional on the fixed sources and response functionals. Thus, the band recovered many correct claims withheld by the ball while achieving empirical family coverage near 95%.
For these two behaviors in the original panel, all 403 band-supported local signs also remained positive at the means of 24,570 unused source records.
We also checked whether the recall gain persisted after adjusting both methods toward the same coverage target. Using known finite-source truth, we adjusted their widths to target 95% simultaneous coverage of the response functionals. On held-out reconstructions at n=1000, the band retained recall gains of 18.9 percentage points under ordinary sampling and 9.2 under stratified sampling over the adjusted ball-derived bounds. Achieved coverages were close but not identical; this truth-based adjustment is a diagnostic, not a procedure for an unknown population. These results suggest that the support gain is not solely due to lower coverage.
At n=128, band coverage was 93.8–93.9%, so small-sample calibration remains a limitation. Bootstrap settings differ in the number of Gaussian multiplier draws (1,999 in validation versus 19,999 in the public-panel comparison) and covariance denominators (sample size minus one versus sample size). These finite-source checks do not establish a 95% guarantee for the original 788 supports or an unknown generating population.